glean-review

name	glean-review
description	Review delivered glean fragments interactively - star, archive, or trash each one, capture a short note on why, feeds signals into feedback.jsonl
allowed-tools	Bash, Read, Edit
user-invocable	true

glean-review

Walk the user through their inbox of newly delivered story fragments. Capture star/archive/trash decisions plus an optional note. This feeds future auto-tuning.

Steps

1. Fetch the review queue. The worker prints a JSON array of inbox fragments (status: inbox, sorted by captured desc):

   PYTHONPATH="${CLAUDE_PLUGIN_ROOT}/scripts" python3 -m glean.worker --review
   

   Each item should include at minimum: fragment_id, path (absolute), title, project, captured, draft_angle (from the ## Thread section), and an excerpt (first paragraph of ## What happened or similar).

2. If the array is empty, print Inbox is empty. Nothing to review. and exit.

3. Initialize tallies: starred=0, archived=0, trashed=0, skipped=0. Track remaining = len(items).

4. For each fragment, display a compact card:

   [i/N]  {title}
   project:  {project}          captured: {captured}
   thread:   {draft_angle}
   
   {excerpt}
   
   path: {path}
   

5. Ask the user with AskUserQuestion:

   • question: "What do you want to do with this fragment?"
   • options: Star, Archive, Trash, Skip, Quit

6. If the action is Star or Trash, ask a follow-up via AskUserQuestion for a one-line note on WHY (especially important for trash). Allow an empty note. Skip the note prompt for Archive and Skip unless the user wants to add one.

7. Apply the decision:

   a. Update the fragment file's frontmatter status field using Edit:

   • Star -> status: starred
   • Archive -> status: archived
   • Trash -> status: trashed
   • Skip -> leave as inbox, do not write feedback

   b. Append a feedback entry to ${CLAUDE_PLUGIN_DATA}/feedback.jsonl (one line, JSON):

   {"fragment_id": "...", "action": "star|archive|trash", "note": "...", "timestamp": "<ISO-8601 UTC>"}
   

   Prefer calling the worker if it exposes a feedback-writing subcommand; otherwise append directly with a shell heredoc. NOTE: CONTRACTS.md does not specify a dedicated feedback CLI flag - if none exists, append via shell and say so clearly in a comment to the user.

8. After each decision, print the running tally and remaining count:

   tally: * {starred} starred  | = {archived} archived  | x {trashed} trashed  | - {skipped} skipped  | remaining: {remaining}
   

9. If the user picks Quit, stop immediately, print a final summary, and exit. Do not process remaining items.

10. After the loop ends (all reviewed or user quit), print a final summary:

    glean-review session summary
    ----------------------------
    reviewed: N   starred: N   archived: N   trashed: N   skipped: N
    feedback written to: ${CLAUDE_PLUGIN_DATA}/feedback.jsonl
    

Notes

• Do not batch-apply. One fragment at a time, wait for user input.
• Never delete fragment files. Trash only changes the frontmatter status.
• If Edit fails on a fragment (file moved, frontmatter malformed), surface the error, do NOT write a feedback entry for that item, and continue to the next.
• Timestamps must be UTC ISO-8601 with Z suffix.